Identity
The actor, customer, workspace, and credential must resolve.
Intelligence never becomes authority by accident.
Omnicort separates what an employee can understand from what it can do. Identity, tenant, consent, evidence, permission, approval, readiness, and audit remain explicit control boundaries around every meaningful action.
Fail-closed action governance
Follow one request through the complete control plane.
Change the scenario to see why the same employee may act, ask the owner, or stop.
ACTION GATE TRACE
See why an employee may act, ask, or stop.
Reserve an approved calendar slot for a verified customer.
- 01Identitypass
- 02Tenantpass
- 03Permissionpass
- 04Approvalpass
- 05Auditpass
Missing proof is not permission
Every external effect must cross the gates that apply to it.
A fast AI answer cannot bypass a slower business obligation. When identity, consent, scope, evidence, approval, wallet, connector readiness, or duplicate protection is missing, the system should fail closed and show the owner what is needed.
Consent
The customer and channel posture must allow this contact or data use.
Scope
The employee needs exact knowledge, channel, tool, and action permission.
Evidence
The claim or decision needs an approved source and current context.
Approval
High-risk or out-of-authority work must have an exact owner decision.
Readiness
Connector, wallet, policy, duplicate, and operational checks must pass.
Tenant isolation
Every business, employee, customer, and record stays inside its permitted boundary.
Workspace membership, server-side authorization, row-level security, customer-private scope, and service-role checks work together. A valid login is not automatically valid access to another tenant, employee, or customer.
Business ABrainEmployeesCustomers
Cross-tenant read denied
Business BBrainEmployeesCustomers
Action firewall
Every meaningful decision resolves to a customer-safe outcome.
The employee does not receive one global autonomy switch. Policy and context resolve each decision into Act, Ask now, Ask later, Brief, or Block.
Act
Authorized, reversible work inside current policy.
Ask now
A customer or high-value decision is waiting.
Ask later
Owner judgment matters, but interruption can wait.
Brief
Useful completed work or a signal that deserves visibility.
Block
Identity, consent, evidence, cost, policy, or safety is not ready.
Memory Trust Firewall
Untrusted content cannot silently become company truth.
The same system that protects actions must protect learning. Otherwise one malicious document, mistaken customer, or one-off exception could change every future employee.
Untrusted by default
Documents, websites, transcripts, messages, connector payloads, and tool output begin as evidence.
Injection detection
Instructions hidden inside external content cannot become system authority.
Private-data boundary
Customer-private facts stay with the exact permitted customer timeline.
Conflict quarantine
Contradictory facts remain visible with both sources and wait for resolution.
Source reputation
Authority, recency, approval, and evidence quality shape inclusion.
Governed promotion
Reusable memory requires scope, provenance, disposition, evaluation, and version lineage.
Decision lineage
An important action should be explainable from evidence to outcome.
Audit is not a pile of raw logs. Owners need a human-readable chain showing what was requested, which source and policy applied, who approved it, what executed, what happened, and which version can be restored.
- 01
Request
Customer asks to move a paid booking
- 02
Evidence
Verified identity + booking record + policy v12
- 03
Decision
Reschedule allowed inside 24-hour rule
- 04
Approval
No owner approval required for this scope
- 05
Execution
Calendar event changed once; idempotency recorded
- 06
Outcome
Confirmation delivered; customer timeline updated
Owner controls
Every permission needs an obvious way back.
Pause
Return an employee or channel to a stopped state immediately.
Revoke
Remove a tool, connector, knowledge, action, or member permission.
Narrow
Reduce authority by employee, action, channel, risk, or customer scope.
Disconnect
Separate a workspace from an external app without broad hidden access.
Restore
Roll back to the exact earlier approved Brain or employee version.
Inspect
Read the source, policy, approval, execution, and outcome behind a decision.
Responsible AI boundaries
What Omnicort should never do silently.
- Turn raw customer messages or connector payloads into approved company truth
- Share raw customer or business data across tenants
- Publish pricing, refund, legal, medical, financial, or public-claim changes without review
- Give every employee every tool or every piece of company knowledge
- Turn one owner approval into broader permission than the exact action requested
- Describe proof-only or disabled capabilities as live customer behavior
Report a security concern
Tell us privately and include enough detail to reproduce the issue safely.
Email info@omnicortai.com with the affected surface, observed behavior, impact, and reproduction steps. Do not include customer secrets or attempt to access data that is not yours.
This page describes product principles and current operating posture. It does not claim a certification that Omnicort has not explicitly published.